Does llms.txt work? What the data shows

No, in practice it does not work to increase AI citation or Google ranking, at least based on what the data shows so far. Google has publicly confirmed it does not use the file, no major LLM has declared official support, and the few large-scale studies that measured actual access to the file found almost no use. That said, implementing it takes minutes and has no known downside, so the "is it worth it" answer depends less on effectiveness and more on your type of site.

llms.txt is a plain-text Markdown file, published at the root of the domain (yoursite.com/llms.txt), proposed in 2024 by Jeremy Howard as a way to list, in a structured format, the most important pages for a language model to read. The idea was to work as a sitemap written for AI instead of for traditional search engines.

In July 2025, Gary Illyes, from Google's Search team, confirmed that Search does not use llms.txt and that there are no plans to start. John Mueller reinforced the position by comparing the file to the old keywords meta tag, a standard any site could declare about itself with no verification, which is why it could never be trusted as a ranking signal. Mueller repeated that position in June 2026, so this is not an isolated or outdated comment (seohappyhour.com).

(Sources: oxigenweb.com.br and usecontentor.com)

Five independent studies, different periods, different methodologies, and all of them converge on the same pattern: the file exists on a small fraction of sites, and even when it exists, almost no one reads it. Five sources pointing in the same direction is a sign of consistency, not an isolated sample.

In practice: if your goal is AI citation in the short term, llms.txt is not the place to invest time. The expected return, based on this data, is close to zero.

It is worth noting: this question keeps coming back in SEO forums with unusual frequency for a technical feature with such limited impact. "Has anyone used llms.txt?", "where are we with this?", "is it worth it" show up again every few months, usually with no clear resolution in the thread. There is a simple explanation: implementation is cheap, and the marketing discourse around GEO keeps pushing the file as "essential," so professionals keep testing it individually without sharing results, and the cycle of doubt starts over.

Part of the recurring confusion comes from two files with different purposes being treated as synonyms. llms.txt is meant as a lean index, with links and short descriptions, closer to a table of contents. llms-full.txt is a separate proposal that tries to embed the full content of the listed pages inside the file itself, meant to reduce the need for the agent to follow each link individually. None of the studies cited above measured adoption or reading of llms-full.txt separately, so there is no specific data on it here. In my assessment, there is no reason to expect a different result from the lean llms.txt: if the problem is that the short file already goes unread, a heavier file to keep in sync tends to cost more work for the same return. If you are going to implement one, start with the lean llms.txt.

There are narrower use cases where the file makes sense, independent of search citation:

• Technical documentation and developer products. The most cited real implementations come from companies like Cloudflare, Mintlify, and Anthropic, precisely in this profile, where AI coding tools (assistants inside IDEs) can consume the file directly during development, outside the search context.
• Products with a stable structure, where keeping a second list of pages updated does not turn into duplicate work with every publication.
• Low opportunity cost. If your CMS or SEO plugin already generates the file automatically (some WordPress plugins do this), there is no strong reason not to have it, even with no expectation of results.

Where the effort does not pay off: content blogs chasing citation in AI answers, e-commerce with a large catalog (keeping the list in sync turns into constant manual work), and any site that changes structure frequently.

Recommended minimal structure, in Markdown:

# Company Name

> One sentence describing what the company does.

## Main pages

- [Page name](https://yoursite.com/page): short description of what is there
- [Another page](https://yoursite.com/other): short description

## Documentation

- [Getting started guide](https://yoursite.com/docs/getting-started): short description

After publishing, check your logs to see if anything beyond your own test is actually accessing the file:

grep "llms.txt" access.log | wc -l

If the number after a few weeks is close to zero, that confirms the pattern from the studies above for your specific site, and it is a good reason not to invest more time in this than the minimum necessary.

// real case: in none of the cases I have tracked did llms.txt prove effective. It still looks like a weak hypothesis, but a cheap one to test.

If you already published an llms.txt, the logs over the next few weeks will show whether anyone reads it, and you can compare them with the studies above. If you have not, it can wait: what matters more is making sure the content itself is accessible, as the technical checklist for AI crawlers shows. This article is part of the pillar GEO in practice: how to get crawled, understood, and cited by AI search engines.